Task-guided Graph Augmentation for Fraud Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing graph mining algorithms for financial transaction networks suffer from high generalization errors due to data sparsity, noisiness, and dynamics, which affects accuracy and robustness in tasks like anti-money laundering and fraud detection.
Innovation Solution
The TGEditor framework provides a task-guided temporal graph augmentation method that preserves the temporal and topological distribution of financial transaction networks, using a multi-resolution temporal generative model and adversarial training to enhance performance in node classification and fraud detection.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional graph mining algorithms are used on financial transaction networks, then the algorithms can process the data, but they suffer from high generalization errors due to data sparsity, noisiness, and dynamics
Solution Approach 1:
The patent applies preliminary action by performing data augmentation before the main graph mining task. The TGEditor framework generates augmented temporal graphs by adding missing edges and nodes based on learned patterns from historical data, thereby proactively addressing data sparsity before the classification algorithms are applied. This preliminary enrichment of the graph structure improves downstream task performance without requiring additional labeled data.
Solution Approach 2:
The patent employs copying by creating synthetic copies of existing graph structures through temporal graph augmentation. The framework generates multiple augmented versions of the original financial transaction network by replicating and transforming existing patterns, relationships, and temporal sequences. These copied structures serve as additional training samples that mimic real data distributions while addressing sparsity issues.
2Reliability
If traditional graph mining algorithms are used on financial transaction networks, then the algorithms can process the data, but they suffer from high generalization errors due to data noisiness
Solution Approach 1:
The patent applies the taking out principle by extracting and separating noisy or incorrect relationships from the graph structure. The TGEditor framework identifies and removes spurious edges and invalid temporal relationships during the augmentation process, thereby eliminating harmful noise while preserving genuine transaction patterns. This extraction of harmful elements cleanses the data before it is used for training.
Solution Approach 2:
The patent converts harm into benefit by transforming noisy and sparse financial transaction data into useful augmented training samples. The framework learns from the existing noisy data to generate augmented versions that highlight genuine patterns while filtering out noise. The very sparsity and noise that cause problems are leveraged to create more robust training examples that improve generalization.
3Reliability
If traditional graph mining algorithms are used on financial transaction networks, then the algorithms can process the data, but they suffer from high generalization errors due to data dynamics
Solution Approach 1:
The patent applies dynamics by creating a temporal graph augmentation framework that explicitly models and preserves the dynamic evolution of financial transaction networks. The TGEditor framework generates augmented graphs that maintain temporal consistency and evolve according to learned dynamics patterns, allowing the system to adapt to changing transaction behaviors over time while preserving the inherent dynamics of the data.
Solution Approach 2:
The patent employs parameter changes by transforming the static graph representation into dynamic temporal graphs with evolving parameters. The framework adjusts graph parameters such as edge weights, node attributes, and temporal relationships based on learned patterns from historical dynamics. This allows the augmented data to reflect parameter changes over time, improving the model's ability to handle dynamic financial transaction patterns.
4Reliability
If graph augmentation is performed to address data sparsity, then the performance of prediction models is improved, but the complexity of the system increases
Solution Approach 1:
The patent applies universality by designing the TGEditor framework as a multi-functional system that performs multiple tasks through a unified architecture. The same temporal graph augmentation module serves both data sparsity mitigation and noise filtering functions, while the learned temporal patterns are reused across different financial transaction analysis tasks. This multi-functionality reduces overall system complexity compared to having separate specialized modules for each problem.
Data Source
AI summary
A computer-implemented method for task-guided graph augmentation and editing includes receiving an input graph in an observed financial transaction network. A data augmentation function is learned, where the data augmentation function maintains a true data distribution of the input graph. An augmented financial transaction network is generated that enhances performance of a downstream task and preserves topological and temporal properties of the input graph.